Public service application completeness desk

The business situation

Illustrative reference architecture. Validate requirements for your organization.

Illustrative architecture: A municipal agency receives applications for a community assistance program through online and paper channels. Caseworkers must identify missing documents, locate applicable published rules, and communicate next steps. Agents support intake and evidence organization; authorized officials retain eligibility, benefit, and appeal decisions.

Problem

Applicants use different languages and document formats, and records may be incomplete for legitimate reasons. Treating an absent document as an adverse fact can unfairly constrain access. Rules also vary by program version and application date.

Why a multi-agent system?

Document intake, rule retrieval, evidence completeness, and communication specialists separate extraction from judgment. A case workflow preserves applicants' submitted evidence, tracks unanswered questions, and routes decisions to accountable officials.

Intended outcome

The intended result is a reviewable operational decision with source evidence and explicit ownership. Measure caseworker preparation and applicant comprehension alongside correction burden. The workflow should make public-service processes more legible while avoiding automated determinations or unsupported claims of regulatory compliance.

Specialist agent team

Document intake agent

Extract submitted fields with page references and uncertainty labels.

Tools

• Document OCR
• approved translation service

Authority boundary

Cannot alter original submissions or infer missing personal circumstances.

Program rule agent

Retrieve published requirements effective for the application date.

Tools

• Program policy index
• effective-date registry

Authority boundary

Cannot create eligibility rules or decide exceptions.

Completeness reviewer

Compare available evidence with the official application checklist.

Tools

• Checklist service
• case evidence query

Authority boundary

Missing evidence remains a question, not an automatic denial.

Applicant communication agent

Draft clear next-step notices in the requested supported language.

Tools

• Approved notice templates
• case draft API

Authority boundary

Officials approve decisions and notices; the agent cannot send adverse determinations.

Enterprise systems

Case management system

Owns application identifiers, assignments, official decisions, and review history.

Document and policy repositories

Store original submissions and published rule versions with access and retention controls.

Applicant service portal

Supports accessible communication, document requests, and status tracking without exposing internal protected notes.

Architecture diagrams

flowchart LR
    N0["Public service application completeness desk"]
    N1["Document intake agent"]
    N2["Program rule agent"]
    N3["Completeness reviewer"]
    N4["Applicant communication agent"]
    N5["Evidence ledger"]
    N6["Human decision"]
    N7["Authorized adapter"]
N0 --> N1
    N0 --> N2
    N1 --> N3
    N2 --> N3
    N3 --> N4
    N4 --> N5
    N5 --> N6
    N6 --> N7

sequenceDiagram
    participant O as "Case orchestrator"
    participant S as "Case management system"
    participant H as "Accountable reviewer"
O->>S: Read application receipt and documents
    S-->>O: Ambiguous date extraction
    O->>H: Request source-page verification
    H-->>O: Confirm application date
    O->>S: Select applicable published checklist
    O->>H: Present corrected completeness packet

Runtime flow

1. 1

   Register application

   Deduplicate submission attempts and preserve receipt time, channel, language preference, and the applicant's original documents.

2. 2

   Authorize case access

   Resolve program, staff assignment, and purpose before retrieving supporting records from other agency systems.

3. 3

   Extract evidence carefully

   Capture document fields with page references; mark illegible or uncertain text for staff review rather than completing it speculatively.

4. 4

   Check published requirements

   Apply the correct checklist version and distinguish missing, unavailable, and disputed evidence in the case packet.

5. 5

   Review applicant notice

   Present officials with cited requirements and a draft request for clarification that explains the available next steps.

6. 6

   Record accountable action

   Save the approved notice or decision through the case system and retain appeal or correction routes under existing agency procedures.

Failure recovery

Security & decision boundaries

• Keep eligibility, benefit amounts, and adverse decisions under authorized official control.
• Never convert missing documents into negative factual assertions about an applicant.
• Provide inspectable source references and correction routes for extracted information.
• Limit cross-agency retrieval by program purpose and approved access agreements.
• Evaluate communication accessibility and language quality with affected user groups.

Operational metrics

Measurement definitions, not promised improvements.

Intake preparation time

Time from received application to caseworker-reviewable evidence and outstanding questions.

Extraction correction rate

Material extracted fields corrected by staff divided by material fields reviewed.

Rule applicability accuracy

Audited checklists using the correct program and date version divided by audited checklists.

Notice comprehension

Applicants correctly identifying required next steps in a defined usability evaluation divided by participants evaluated.

Architecture tradeoffs

Rollout plan

Synthetic application evaluation

Staff and accessibility reviewers validate missing-evidence treatment, language quality, and date-sensitive rules.

Shadow intake pilot

One program compares draft packets with caseworker practice and tests cross-agency access restrictions.

Approved notice assistance

Controlled notice drafts launch after correction, rule-transition, and appeal-route visibility checks pass.

Connected concepts

• 21 — Authorization-Aware RAG
• 23 — Attribute-Based Access Control (ABAC)
• 25 — Data PII Masking
• 28 — Deterministic Data Grounding
• 32 — Model Bias Auditing
• 38 — Data Lineage Tracking
• 44 — Human-in-the-Loop (HITL)

Primary references

• NIST AI Risk Management Framework
• NIST SP 800-207: Zero Trust Architecture
• Temporal: Workflow execution